Struct CpuEncoder 

pub struct CpuEncoder { /* private fields */ }

The CPU encoder: loads a checkpoint directory (config.json + model.safetensors) and embeds ragged token batches. Deterministic per input (fixed reduction orders), so it serves as the GPU kernels’ tolerance-gated oracle.

Implementations

impl CpuEncoder

pub fn load(dir: &Path) -> Result<Self>

Load from a checkpoint dir. Refuses (loudly) architectures whose CPU arm hasn’t landed — “recognized” and “supported” stay honest.

pub fn load_siglip_text(dir: &Path) -> Result<Self>

Load the SigLIP TEXT tower from a joint checkpoint.

pub fn load_siglip_vision(dir: &Path) -> Result<Self>

Load the SigLIP VISION tower from a joint checkpoint.

pub fn encode_image( &self, pixels: &[f32], image_size: usize, patch_size: usize, ) -> Result<Vec<f32>>

Embed ONE image’s normalized pixels ([3 · image² ] CHW, already resized/rescaled/ normalized): im2row patch embedding + positions → transformer → post-LN → MAP head. Vision tower only.

pub fn config(&self) -> &EncoderConfig

The parsed configuration.

pub fn encode(&self, batch: &EncBatch) -> Result<EmbedOut>

Embed a ragged batch. Sequence length (plus any learned-position offset) must fit max_pos; empty sequences are refused (a mean over zero rows has no value).

pub fn forward_hidden(&self, batch: &EncBatch) -> Result<Vec<f32>>

The raw pre-pooling [total_tokens, hidden] hidden-state stream (row-major, ragged- concatenated) — the encoder’s last_hidden_state. Every pooling mode is a reduction of this: Cls L2-normalizes each sequence’s FIRST row, LastToken its last, Mean the row-average. Exposed so parity gates and callers can inspect per-token states directly. Runs inside the encoder-owned rayon pool.